Prediction of the histopathological tumor type of newly diagnosed liver lesions from standard abdominal computer tomography with a machine-learning classifier based on convolutional neural networks

Abstract Background and objectives: Liver lesions are a relatively common incidental finding in computer tomography (CT) of the abdomen. The current gold standard is liver biopsy, which has the downside of respecting only a small part of the total lesion volume. Furthermore, this invasive method carries interventional risks like bleeding or infection. Therefore, an image-based biomarker would be highly desirable. Conventional “radiomics” methods have often been utilized for similar problems, but the results are often not reproducible. This is mainly due to sampling errors and interobserver variability, but also the seemingly complex nature of the problem. We present a new approach that implements cutting-edge research in machine learning which is nevertheless cheap and easily applicable in a routine clinical setting. To achieve this, we use convolutional neural networks (CNN) to predict the histopathological findings from liver lesions from preoperative liver CT. Methods: After splitting the study population into a training and test set we trained a CNN to predict the histopathological tumor type from CT data. Results: The developed CNN workflow is able to predict liver tumor histology from routine CT images. We also evaluated in how far transfer learning and data augmentation can help in solving this problem and implemented the developed workflow in a clinical routine setting. Conclusion: We propose a robust semiautomatic end-to-end classification workflow for the prediction of the histopathological type of tumor lesions based on abdominal CT and a deep convolutional neural network model. In our cohort, the model shows reliable and accurate results even with limited computational resources.